How can I avoid vendor lock-in when using AI coding agents?
Summary
- Lock-in with coding agents happens in layers: the model provider, the agent harness, the tools and workflows wired into one agent, and the usage data sitting in one vendor's console. Keep each layer swappable.
- Route model calls through a provider-independent gateway, so switching models or providers is a configuration change, not a migration.
- Keep tools on the open MCP standard and team workflows in shared skills, so they follow developers to whichever agent they use.
- Keep usage, cost and trace data in your own tables, and benchmark agents on your own code, so the decision to switch rests on evidence.
- On Databricks, Unity Gateway (GA) connects several coding agents to the same governed models and MCP servers, and Omnigent (Beta on Databricks, open source and in alpha) adds a common layer over agent harnesses.
How can I avoid vendor lock-in when using AI coding agents?
Separate the agent from everything around it. Put a governed gateway between your coding agents and model providers, keep tools and team workflows in shared formats any agent can use, store usage and trace data in your own tables, and support more than one agent from the start. When a better model or harness appears, you change a setting instead of rebuilding access, policy and integrations.
The pressure to switch is constant. Databricks' AI product and engineering team, in its Unity Gateway CLI announcement, describes the trade-off: standardize on one provider and you risk missing better models or lower costs elsewhere, but support many agents and providers without a common layer and access, budgets, security policies and usage logs scatter across separate systems.
Where lock-in hides
| Layer | How lock-in happens | How to stay portable |
|---|---|---|
| Model provider | The agent only talks to one provider's API and account | A gateway with provider-independent APIs, several providers and open-weight models |
| Agent harness | Workflows, prompts and settings only work in one tool | Support several agents, or a meta-harness that sits above them |
| Tools and data access | Integrations are rebuilt for every agent | MCP servers registered once and governed centrally |
| Team workflows | Instructions live in one agent's private format | Shared skills that every supported agent can load |
| Usage and traces | Adoption and spend data stay in each vendor's console | Logs and traces in tables you own and can query |
| Evaluation | Choices rest on launch claims and public leaderboards | A private benchmark built from your own code |
How Databricks approaches agent portability
The authors of Databricks' coding benchmark post put the principle plainly: "we've always been wary of lock-in, not just to vendors, but to assumptions that make teams less flexible over time." The same benchmark found that the best quality for a given cost came from a mix of providers and open models, and that the harness a model runs in changes both cost and quality. Databricks builds for that mix:
Many agents, one set of models. Unity Gateway, which is generally available, connects coding agents such as Claude Code, Codex and OpenCode to models from multiple providers, MCP tools and shared skills, and the Unity Gateway CLI (ug) launches them with one command. The docs' capability matrix shows exactly which of those each supported agent can use.
Models you can swap. Databricks hosts frontier proprietary and open-weight models, supports open models with coding agents, and connects external providers you already pay for through a governed model provider service. For applications, the unified APIs translate each request to the model's native format, so teams switch models across providers without changing client code.
Tools and skills that travel. MCP, the open Model Context Protocol, connects agents to data and services; MCP servers are governed as Unity Catalog securables, and ug mcp add adds them to an agent. ug skills add brings shared skills into supported agents; skills require a Beta feature your admin enables.
Data you keep. Every gateway request lands in the system.ai_gateway.usage system table, with cost and audit data alongside, so adoption and spend analysis doesn't depend on any one agent vendor's console.
A layer above harnesses. Omnigent is an open-source meta-harness under the Apache 2.0 license; its open-source repo marks it as alpha, and the managed version on Databricks is in Beta. The docs describe it as a common layer over Claude Code, Codex, Cursor, Pi and agents you write yourself: define an agent in a short YAML file, then change its harness or model while tools, prompts, skills and policies stay the same.
Use cases
- A platform team supporting several agents. Developers launch their preferred agent through
ug, under the same models, MCP servers and budgets. - A model switch without a migration. When a cheaper or better model proves itself on your benchmark, admins grant access to its model service; developers pick it up in the agents they already use.
- An exit plan for a harness. Teams that define agents in Omnigent can move a workflow to a different harness by changing configuration, not rewriting the agent.
FAQs
Doesn't a gateway create its own lock-in?
We believe the risk is lower when the gateway speaks open interfaces: Unity Gateway uses OpenResponses- and OpenAI-compatible APIs and the open MCP protocol, and its usage logs land in Unity Catalog tables you can query.
Can I use open models with coding agents?
Yes. Unity Gateway supports open models with coding agents, and the supported agents page lists which agents can use them.
Do developers have to change how they work?
No. ug opens each agent with the same interface and workflow developers already know, connected to Unity Gateway.
How do I know when to switch agents or models?
Benchmark them on your own code. Databricks built a benchmark from its merged pull requests and plans to run every new agent and harness through it.
The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.